A deep-learning framework for forecasting renewable demands using variational auto-encoder and bidirectional long short-term memory

Abstract

This paper aims to suggest a deep-learning framework for large-scale renewable demands using 1) variational auto-encoder (VAE) algorithm to generate a number of feasible samples and 2) bidirectional long short-term memory (Bi-LSTM) networks to construct a demand forecasting model for renewable electricity. Pre-and post-processing of raw data is carried out to formulate conversion factors that are explicitly involved with data sampling and data distribution. Other machine learning methods such as gated recurrent unit, long short-term memories, artificial neural network, deep neural network, support vector regression, and auto-regressive integrated moving average are considered together to compare with the proposed forecasting model and all of which are applied for a case study of South Korea aiming for a greener economy. Performance evaluation metrics incorporating root mean square error (RMSE), mean squared error (MAE), mean absolute percentage error (MAPE), and R-squared (R2) results in that the VAE-Bi-LSTM-based forecasting model outperforms other models, indicating that values of the RMSE, the MAE, and the MAPE decrease by 33.7%, 41.4% and 39% less on average and the R2 score increases by 3.5% more on average. Also, results from information criteria show that the VAE-Bi-LSTM based forecasting model is the most optimal network among RNN-based models. It is expected that this study would contribute to a variety of large-scale energy policies.